TLDR Hardware 2026-08-31

CXMT is suing the Department of Defense to overturn its January 2025 classification as a Chinese military-linked company, arguing it has no ties to the PLA and that the designation lacks evidence and violated due process. The company says the Pentagon's original case leaned on a distributor's unauthorized "military grade" ad, later admitted erroneous, and that its DRAM actually conforms to civilian JEDEC commercial standards, lacking the temperature, packaging, testing, and oversight requirements imposed on genuine Chinese military-grade semiconductors.

Andreessen Horowitz raised $1.1 billion for a new "Machine Age" fund, a notable pivot for a firm best known for backing software, explicitly aimed at chips, memory, interconnects, data centers, cooling, and power-efficient edge devices rather than another model or app layer. The firm frames the bet around needing faster, more efficient systems, cheaper high-bandwidth memory, and better interconnects between nodes to keep the AI buildout from being bottlenecked by physical infrastructure, alongside broader materials, electrical, and real estate build-out to support it all.

Meta is piloting robots from Watney Robotics, Kinova, and ABB to handle physical data center work, evaluating whether a Kinova Gen3 robotic arm can power-cycle servers, testing a separate robot for swapping networking cables, and running ABB units mounted on wheeled platforms to reseat components at its newer Prometheus campus in New Albany, Ohio. It builds on simpler automation already in place, including self-driving tugger robots that move heavy server racks and wheeled inventory robots that scan equipment. One Meta worker estimates a successful cable-swapping robot alone could replace up to 80% of that specific role's workload.

NVHBM is a design that moves NVIDIA's custom memory controller from the accelerator into the high-bandwidth-memory base die. The change frees compute-die area and lets multiple memory providers build compatible stacks, with Amazon's Annapurna Labs as the first collaborator.

Reimagine Robotics, founded by former Google DeepMind applied-robotics lead Jonathan Scholz and other ex-DeepMind engineers, is building robots that factory workers can post-train directly through demonstrations and physical corrections. In one industrial deployment involving recovery of critical materials from used hard drives, Reimagine says the approach reduced the time needed to teach a new task from one day to 10 minutes. The architecture shifts adaptation from centralized robotics engineers toward workers operating the robots in production environments.

Pushing gallium nitride to higher voltages usually causes charge to pile up in one spot when the transistor switches off, triggering electrical breakdown. EPFL's POWERlab fixed this by engineering an additional layer that produces a sheet of mobile holes nearly matching the electron sheet GaN naturally generates, letting the two charges cancel out so the electric field spreads evenly instead of spiking, all from the material's own polarization rather than temperature-sensitive doping. The result: diodes breaking down above 3.9 kV and transistors sustaining 3.4-3.5 kV, more than 5x the rating of many commercial GaN devices, holding up at 125°C with under 15% resistance drift during repeated high-voltage cycling.

Lambda raised roughly $1 billion in private short-dated debt, arranged by JPMorgan, to buy Nvidia GPUs that Microsoft will then lease, borrowing against Microsoft's willingness to keep paying rather than its own revenue, letting Microsoft get the compute without carrying the debt itself. It's Lambda's second such raise this month, following a $917 million loan two weeks ago secured against a contract with Nvidia, which is simultaneously its investor, supplier, and customer, and the company is reportedly in talks for up to $3 billion more ahead of a possible IPO next year

Elon Musk has confirmed that a secretive SpaceX foundry in Bastrop, Texas is meant to cast turbine blades in-house, tackling a manufacturing bottleneck he says currently limits how fast natural gas turbines can come online for AI data centers, up to 18 months faster if it works. The blades sit in a turbine's hottest section at 3,000-3,600°F, roughly 800°F above the melting point of their own alloy, survivable only through internal cooling channels, thermal coatings, and a single-crystal casting process grown slowly in a vacuum furnace, a technique just four companies worldwide have mastered at industrial scale, and all are currently sold out.

Pollen Robotics and Hugging Face opened pre-orders for Microduck, a 25 cm open-source biped whose policies are trained in MuJoCo and then deployed to the physical robot.

ChangXin Memory Technologies has started mass production of LPDDR6, its most advanced mobile DRAM yet, and confirmed it will supply the chips for Xiaomi's upcoming 18 Fold, launching in September, pairing the memory with Xiaomi's own 3nm Xring O3 processor

SK hynix held a ceremonial groundbreaking for its first US high-bandwidth memory production hub, a $4 billion advanced packaging and R&D facility in Indiana's Purdue Research Park first announced back in April 2024.

Ken Shirriff takes apart the core memory module from a 1980 Spacelab computer.

At Hot Chips 2026, Chips and Cheese sat down with IBM's Christian Zoellin and Christian Jacobi to dig into the engineering behind IBM's next-gen mainframe chip, the first to run both z/Architecture and Arm instructions natively on one core, plus the shift to HBM in its upcoming Spyre AI accelerator.